BrainOutside
beta
Your brain, kept outside your head.
Where your agents can read it.
One git repo of markdown is your brain: your takes, your stories, your beliefs, your voice. BrainOutside is two ways to use it: read it with Claude Code on your laptop, or serve it to every agent you run.
Your brain is a git repo of markdown
Not a vector database. Not an app's proprietary silo. Plain files in a normal repo, because a brain you cannot read is a brain you cannot trust.
You can read it
Open it in any editor. git log it. Fix a note by editing a file.
You can leave
It is your repo. Delete the server and the brain is still there, intact and useful.
It has history
Change your mind and the old note is superseded, not deleted. A record of how your thinking moved, not just where it landed.
Two ways to run a brain
Local: VS Code + Claude Code
free · private · zero infra- Clone the template. The contract, the note templates, and two skills come with it.
- Feed it. Say "feed this into my brain" with a link or a thought. The agent proposes notes, and nothing is written until you approve.
- Ask it. "Using my brain, what's my take on…" Retrieval through a lens, answers in your voice.
Self-hosted: the online head
your VPS · every agent- MCP + REST. Point Claude Code at your own mind. Or any agent, any pipeline, anywhere.
- Tiers enforced server-side. Public, agents-only, private. Each API key sees only its tier, structurally.
- The same human gate. Feeds land in an approval queue; approval is one commit back to your repo, authored by the server so its writes never look like yours.
And the two ways compose.
The local repo is the repo the server clones. Start local today; add the server when you want it. Nothing to migrate. Same brain, two heads.
Where this came from
This project grew out of Andrej Karpathy's llm-wiki and the wave of markdown knowledge bases around it. The core insight there is right: plain markdown plus a coding agent beats RAG for personal knowledge. We surveyed the landscape of memory and knowledge-base tools, open source and commercial, and built where they stopped:
- A wiki holds knowledge, not you. Identity, voice and beliefs are first-class here, so agents can write as you, not just recall.
- No safe way to grow. A self-maintaining wiki fills up with unreviewed extractions. Here every write is proposed by an agent and approved by you.
- Recall isn't the point. Creation is. Lenses and context packs turn the brain into a writing instrument: replies, posts, scripts, in your own voice.
- A local folder serves one tool on one machine. The server is the brain's online head: self-hosted, private, readable by every agent you run.
The llm-wiki was not the only influence. If you recognize your project in this lineage and want a link here, open an issue.
What the server adds
Nothing enters without you
Agents propose; you approve in a diff view; approval is one commit, under the server's own name.
Tiers that are real
Each tier is a materialized snapshot. A reader agent physically cannot read above its tier.
MCP + REST
One endpoint registry serves both, plus self-generated API docs on your instance.
No API billing required
Runs on a Claude subscription token as a first-class path. An API key works too.
A full ledger
Every token, every read, every SDK run: logged, priced, and on a dashboard.
The mind, visible
Visibility rings, a topic graph, live read activity, and a timeline of every position you revised.
The honest part
Private notes are only as private as your VPS. Tiers are enforced server-side and writes are human-gated, but a self-hosted brain lives in plaintext on a machine you administer. We document the real security posture instead of pretending otherwise. Read it before you feed anything sensitive.